Quality assessment of individual classifications in machine learning and data mining

摘要

Although in the past machine learning algorithms have been successfully used in many problems, their serious practical use is affected by the fact that often they cannot produce reliable and unbiased assessments of their predictions' quality. In last few years, several approaches for estimating reliability or confidence of individual classifiers have emerged, many of them building upon the algorithmic theory of randomness, such as (historically ordered) transduction-based confidence estimation, typicalness-based confidence estimation, and transductive reliability estimation. Unfortunately, they all have weaknesses: either they are tightly bound with particular learning algorithms, or the interpretation of reliability estimations is not always consistent with statistical confidence levels. In the paper we describe typicalness and transductive reliability estimation frameworks and propose a joint approach that compensates the above-mentioned weaknesses by integrating typicalness-based confidence estimation and transductive reliability estimation into a joint confidence machine. The resulting confidence machine produces confidence values in the statistical sense. We perform series of tests with several different machine learning algorithms in several problem domains. We compare our results with that of a proprietary method as well as with kernel density estimation. We show that the proposed method performs as well as proprietary methods and significantly outperforms density estimation methods.